So in AI, there’s this thing called data centers. You probably know about them (maybe you hate them!). But what is a data center? Picture a giant warehouse full of metal racks; each rack packed with specialized chips that crunch the numbers behind every ChatGPT, Claude, or Gemini request.
And so far, more AI demand = more chips = bigger racks = more and bigger buildings.
Well, the big story today is a chip company called Cerebras. The most interesting thing about them is that, instead of cutting a silicon wafer into hundreds of smaller chips, they make one chip the size of a dinner plate. Oh, also, OpenAI is paying them $20B+ for 750 megawatts (MW) of inference compute (the aforementioned AI number-crunching) through 2028.
Well, they’re going public this Thursday, and just upsized the IPO.
Here are the details
- Pricing jumped from $115-125 to $150-160 per share in days.
- Cerebras will now aim to raise $4.8B at a ~$33B valuation.
- Orders came in for 20x the shares available.
- Shares debut on Nasdaq as “CBRS” this Thursday, May 14.
Here’s why that’s a big deal: Nvidia is the most valuable company in the world (well, trading places off and on with Google these days), and Cerebras is a real threat. Why? For inference (the part where an AI generates a reply to your prompt), bigger chips mean less data shuttling between chips, which means faster and cheaper answers.
Nvidia agrees so much that last December, it paid $20B to acqui-hire Groq, a rival chip startup (Groq the company, not Grok Elon’s chatbot, to be clear). And remember last week, when xAI leased its Colossus 1 facility, all 220,000 Nvidia chips, to Anthropic for ~$5B a year? Elon is basically a cloud landlord now, too.
Why should YOU care? Cheaper, faster inference = better AI tools at a lower bill. Also, not financial advice and shared for educational purposes only, but public companies are investable by the public, so y’know, there’s that.
Demand for data centers is everywhere
- A startup is helping OpenAI tailor its models to Cerebras silicon, because Nvidia chips are too scarce to rely on alone.
- SoftBank’s Masayoshi Son is in talks with Macron to put up to $100B into French data centers.
- And Cowboy Space raised $275M to build data centers in orbit; per TechCrunch, the bottleneck is rocket capacity, not engineering.
Why aren’t MORE people freaking out? Because demand for intelligence might be the closest thing the economy has ever seen to demand for energy: virtually limitless.
There are no low-energy, high-income countries on earth; every wealthy economy runs on enormous amounts of electricity. Anthropic’s Dario Amodei argues intelligence is now a basic factor of production, like land, labor, or capital. If he’s right, every economy that wants to grow will have to consume vast quantities of it. Today’s spending starts to look less like a bubble and more like laying the rails for the next industrial revolution.
That unlimited demand also has to live somewhere physical. Stratechery’s Ben Thompson, writing about this exact IPO yesterday, splits inference into two types:
- “Answer inference” is what we use today: chatbots with humans waiting, latency-sensitive, GPU-and-Cerebras-shaped.
- “Agentic inference” is what’s coming: AI doing overnight work with no human watching, where memory matters more than speed.
Cerebras is perfectly timed for the first market. The second one looks different, and can run on slower, cheaper, or even orbital compute… wherever electricity is cheapest.
Editor’s note: This content originally ran in the newsletter of our sister publication, The Neuron. To read more from The Neuron, sign up for its newsletter here.